Self-Directed Memory Management

Let the agent call memory tools itself — guided by hierarchy docs and pressure signals — rather than only a fixed external eviction policy.

Self-directed memory management makes overflow handling part of the agent's function calling loop: the model decides when to move items between agent memory tiers, edit working state, or search durable stores. The harness still owns parsing, validation, and execution; the policy of what to keep in-window is model-generated from current context.

Two prompt ingredients make this workable: a detailed description of the memory hierarchy and each store's utility, and a function schema (with natural-language docs) for access and mutation. Feed execution results — including capacity errors — back into the next turn so the agent can adjust. Memory pressure eviction warnings teach awareness of context limits; without them, autonomous edits often arrive too late.

Use when conversation or task state outlives a single window and a fixed FIFO or summary policy would drop the wrong facts. Pair with function-call chaining for multi-step retrieval, and keep retrieval results page-sized so tools cannot blow the budget the agent is trying to manage. Document QA is the stress case: paginated archival search lets the agent iterate when top-(K) misses gold — but only if the model actually continues paging instead of answering early (OS-inspired agent memory). Encode that obligation as an exhaustive archival search instruction in the persona/system layer, not only in tool docs.

---

Sources

• MemGPT (Packer et al.), §2